QwenLM / Qwen2

Qwen2 is the large language model series developed by Qwen team, Alibaba Cloud.

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

LoRA训练和推理精度问题

LiuuHuanChang opened this issue · comments

请问有人用float16精度训练过LoRA吗,推理用bfloat会不会影响推理效果。当前训练精度用的float16,但推理由于总是OOM所以只能用bf来推,但是现在推理效果上看感觉完全没训练进去,不知道是训练的问题还是推理load模型的问题。

I think that float16 and bfloat16 should take about the same memory. Do you happen to have the difference in memory after loading the model in the two precisions?

bf16和fp16在内存占用上都是16位,正常来说,它俩占用的内存是一样的